Explainable deep learning for time-series medical data predictions

Abstract

In recent years, Deep Learning (DL) has been at the forefront of Artificial Intelligence, driven by the increasing availability of computational resources. However, despite its advancements, DL still faces serious challenges, particularly in high-stakes areas such as healthcare. The complexity and lack of interpretability of DL models lead to serious concern, questioning their reliability and adoption in life-sensitive applications. This paper aims to migrate these issues by leveraging Explainable Artificial Intelligence (XAI) to bridge the gap between model complexity and improve interpretability by providing insights into how DL models make their decisions. This study focuses on improving the interpretability of DL models on vital signs as it is the key parameter for assessing patients' health status and predicting health outcomes. However, this poses a significant challenge for the current XAI methods, which excel in explaining individual predictions but fail to address the sequential dependencies inherent in time-series data, such as vital signs. This paper intends to research and adapt explainability methods that account for the temporal nature of such data to ensure reliable and transparent decision-making in healthcare applications. By highlighting the effect of time-dependent features on model predictions, this study aims to enhance the interpretation of the DL models and trust the AI-driven decisions. These proposed methods aim to promote the adoption of AI in healthcare that uses time-series data by addressing the challenges of transparency and interpretability in life-critical applications.